Extract text inside <answer> tags if think_flag is True. Args: text: Raw model output. Returns: Ext

Randomly sample questions for each domain and format them as output items. Each item contains an empty 'instruction' and an 'input' with th

Return system prompt depending on whether <think>/<answer> tags are enabled.

Collect questions grouped by domain. Returns: dict: {domain_name: [list of questions]}

Compute the BGE-based reward score. Args: solution_str (str): The generated solution string. ground_truth (str): JSON string

Create GRPO training configuration from argparse args.

Extract the actual query text from a solution string. - If `<|im_start|>assistant` exists, strip everything before it. - If `think_f

Load previously reasoned queries if the file exists.

Load and shuffle dataset from parquet file.

Load causal LM model and tokenizer from a given path.

Parse command-line arguments.

Parse command-line arguments.

Parse command-line arguments.

Parse command-line arguments.

Generate model output given a prompt and optional instruction. Args: prompt: User query or input text. instr

Process sampled questions into dataset items. Returns: (list of items, max question length)

Iterate over examples and generate reasoned queries.

Save the current query map to a JSON file.

Setup logging to file and console.

Reward function that computes BGE-based scores for multiple completions. Args: completions (list): List of completions, each a list